4. Comissioning
MiR AI Camera Getting started (en) 09/2019 - v.1.0 ©Copyright 2019: Mobile Industrial Robots A/S.
18
4. Comissioning
This section describes how to train MiR AI Camera to detect certain objects and setup which
actions should be triggered when the camera detects specific objects. There are four main
phases included in the training of MiR AI Camera:
1.
Collection phase
Sample images are collected autonomously by the AI cameras. These are required for
training the AI camera.
2.
Pre-processing phase
Objects in the collected images are detected and framed. Images containing similar
objects are clustered into a groups.
3.
Validation phase
You must label the collected images based on the detected object in the image, and only
select images with accurate framing.
4.
Training phase
Based on the validated images, MiR AI Camera is trained to detect future observations of
the labeled objects.
4.1 Collection phase
For MiR AI Camera to detect specific objects, it must first be trained using a collection of
sample images of those objects. The objects you would like the camera to be able to detect
are referred to as target objects. Each camera autonomously collects the necessary images
over an eight hour period, optimally while the work area is in operation. In other words, do
not leave the camera to collect data after work hours or in a time span where it is not
intended to be used.
It is important that you gather images of objects where they are alone in the
camera's field of view. It is less efficient to train MiR AI Camera when there
are multiple target objects in the images.
When the cameras are set to collection mode, all cameras connected to the fleet are set to
record images whenever they detect motion. It is important during the collection phase that
the desired target objects enter the camera's field of view.
